A neural architecture for fast and robust face detection

García, Christophe; Delakis, M.

doi:10.1109/icpr.2002.1048232

Cited by 53 publications

(22 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Other more recently reported methods [25] are often variations and/or extensions of [21], using Haar-like features or local binary patterns (LBP) or anisotropic Gaussian features filters with AdaBoost type algorithms ( [2], [8], [10], [11], [12], [13], [20], [24]), or employ other techniques like support vector machine, SVM ( [2], [6], [7], [23]), Haar wavelets ( [17]), convolutional neural networks -CNN / ConvNet ( [3]), facial landmarks models ( [26]), or energy based methods ( [14]), while (some) are still using portions of image scanning and preprocessing as in [19] and/or [16]. These methods also demonstrate good (or promising) results, several not only for the frontal-view, but also for the multi-view case.…”

Section: Parallels With Other Methodsmentioning

confidence: 99%

A Pattern Matching Method and Algorithm for Face Detection

Vrejoiu¹

2017

SIC

View full text Add to dashboard Cite

Abstract:The paper presents a simple, original method, using supervised learning and pattern matching, for frontal-view face detection. The raw method and algorithms, their refinements and optimizations, the experimental system, and the obtained results are described incrementally. A pyramid of several simplified representations for faces, with gradual complexity and dimensionality from coarser to more detailed ones, has been defined. These representations were used as a basis for structuring and organizing the knowledge base as a kind of hash tree, as well as to optimize the developed algorithms and their processing time, by applying a sequence of filters with gradual computational complexity in cascade at detection. Appropriate metrics have been defined and used to evaluate the similarity between two finite sets of binary values, of the same dimension, with a minimum of computational effort in these filters. Various thresholds of passage were empirically chosen and adjusted. An experimental system was developed and used for learning and detection tests. Public image databases and private images have been employed, and quite promising results have been obtained. Finally, comparative parallels with some reference methods are discussed.

show abstract

Section: Parallels With Other Methodsmentioning

confidence: 99%

A Pattern Matching Method and Algorithm for Face Detection

Vrejoiu¹

2017

SIC

View full text Add to dashboard Cite

show abstract

“…The similarity between the new class of CoNNs and the existing CoNN architectures [4], [21], [22] is that all of them are based on the same structural concepts of local receptive fields, weight sharing and subsampling. However, their implementations in the new architecture differ markedly.…”

Section: Structural Differences Between Our Network and The Existimentioning

confidence: 99%

“…In our proposed network, the convolutional and subsampling layers are collapsed into one layer, which simplifies the network architecture. Furthermore, the receptive field size in the CoNN used for face detection [21] is different for each convolutional layer, whereas in our approach the same receptive field size is employed throughout the network architecture. In [2] and [21], the connections from the first subsampling layer to the second convolutional layer are manually specified by the user, depending on the specific task to be solved.…”

Section: Structural Differences Between Our Network and The Existimentioning

confidence: 99%

“…However, their implementations in the new architecture differ markedly. The CoNNs used in face detection [21] and face recognition [4] are based on the network architecture developed by LeCun et al [2] for handwritten digit recognition. This network architecture was implemented in such way that each convolutional layer is followed by a subsampling layer.…”

Section: Structural Differences Between Our Network and The Existimentioning

confidence: 99%

See 1 more Smart Citation

Efficient Training Algorithms for a Class of Shunting Inhibitory Convolutional Neural Networks

Tivive

Bouzerdoum

2005

IEEE Trans. Neural Netw.

View full text Add to dashboard Cite

This article presents some efficient training algorithms, based on first-order, second-order, and conjugate gradient optimization methods, for a class of convolutional neural networks (CoNNs), known as shunting inhibitory convolution neural networks. Furthermore, a new hybrid method is proposed, which is derived from the principles of Quickprop, Rprop, SuperSAB, and least squares (LS). Experimental results show that the new hybrid method can perform as well as the Levenberg-Marquardt (LM) algorithm, but at a much lower computational cost and less memory storage. For comparison sake, the visual pattern recognition task of face/nonface discrimination is chosen as a classification problem to evaluate the performance of the training algorithms. Sixteen training algorithms are implemented for the three different variants of the proposed CoNN architecture: binary-, Toeplitz-and fully connected architectures. All implemented algorithms can train the three network architectures successfully, but their convergence speed vary markedly. In particular, the combination of LS with the new hybrid method and LS with the LM method achieve the best convergence rates in terms of number of training epochs. In addition, the classification accuracies of all three architectures are assessed using ten-fold cross validation. The results show that the binary-and Toeplitzconnected architectures outperform slightly the fully connected architecture: the lowest error rates across all training algorithms are 1.95% for Toeplitz-connected, 2.10% for the binary-connected, and 2.20% for the fully connected network. In general, the modified Broyden-Fletcher-Goldfarb-Shanno (BFGS) methods, the three variants of LM algorithm, and the new hybrid/LS method perform consistently well, achieving error rates of less than 3% averaged across all three architectures.

show abstract

“…A great deal of work has been done in this direction ( [21,31,29,10,26,23,24,8,18,14,1] and many others). However the problem is still challenging, because most of the existing methods require thousands of training images of the object.…”

Section: Introductionmentioning

confidence: 99%